acceptodds
Under review as a conference paper at ICLR 2027

Gaussian Mixture Copula Processes for Irregular Time Series

Abstract

We introduce Gaussian Mixture Copula Processes (GMCP), a conditional copula process for irregularly sampled multivariate time series (IMTS) that is expressive and marginalization consistent at the same time. Existing conditional copula processes achieve only one of the two. Gaussian copula processes are consistent by construction but confined to elliptical dependence, whereas attentional copulas such as TACTiS-2 are far more expressive but neither preserve the marginals they are built on nor guarantee that integrating out a target point returns the joint over the remaining ones. GMCP closes this gap by inferring the parameters of a Gaussian Mixture Copula from the context in a way that guarantees consistent marginalization. This yields a non-elliptical copula process with tractable likelihood and consistency guarantees. Our experiments show that copulas equipped with MargFlow, a normalizing-flow model for the univariate marginals fitted in isolation, outperform jointly trained baselines in probabilistic IMTS forecasting. We further provide evidence that GMCP is more expressive than Gaussian copula processes, and on par with the inconsistent TACTiS-2.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Neural Processes, Irregular Multivariate Time Series, Copulas

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rTLxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes a copula process with a dependence modelled by a Gaussian mixture. One of the main sources of novelty is to make the parameters of this mixture dependent on the context through a neural network, parametrised in a way that ensures marginalisation consistency (as explained in lines 195-200).

The proposed model is then benchmarked against several appropriate competitors on synthetic data (5.1) and real-world datasets (5.2).

Critical strengths

The main strengths of the paper are

i) to impose a non-Gaussian (Gaussian mixture) dependence and still achieve marginalisation consistency, which, as the paper highlights, are two features that other methods fail to achieve jointly.

ii) empirical evidence that the GM copula outperforms the mixture of Gaussian copulas (synthetic experiment)

Critical weaknesses

I can identify two weaknesses, of different types.

i) Limited methodological novelty. The proposal mixes a number of concepts known in the community: Gaussian mixture copulas, context-dependent parameters, marginalisation consistency. This does not hinder the strength of the contribution per se, but:

ii) Though the experimental evaluation considers several appropriate benchmark models, I see two sources of concern. First, different experiments are needed to identify the proposed contribution relative to existing methods such as MOSES. For instance, comparing models with the same "capacity", e.g., encoder and comparable parameterisations, but different approaches to modelling marginals and dependence. Second, the empirical validation, though superior on 3 out of 4 datasets, mostly offers marginal improvements (wrt the stated error bars) against TACTiS-2

Questions for authors
  • How can the authors confirm that the performance improvements come from the proposed methodology (copula), and not from differences in the encoder, marginal parameterisation or training procedure? In particular, how can the effect of these contributions relative to other models be isolated?

  • Given how close TACTiS-2 and GMCP are, can you provide a statistical test to verify the proposed model's superiority?

Reviewer_9dapOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Gaussian Mixture Copula Processes (GMCP), a conditional copula process for irregularly sampled multivariate time series. The dependence is a Gaussian mixture copula induced by a latent K-component Gaussian mixture.

This methodology ensures consistent joint marginalization. The copula sits on top of separately trained normalizing-flow marginals (MargFlow) and is trained by numerically inverting the mixture’s marginal CDF, with implicit gradients. On three synthetic processes and four IMTS benchmarks, GMCP attains a better normalized joint negative log-likelihood than jointly trained models and mixtures of Gaussian copulas, and roughly matches an adapted TACTiS-2.

Personally, I believe that this is a well-written manuscript with a broad enough scope. As other reviewers have pointed out, the contribution is somewhat obscured, and the novelty is quite limited. On the existing benchmarks, the proposed methodology achieves very competitive performance.

Critical strengths
  • Exact marginal preservation combined with consistency is well motivated, particularly where univariate forecasts are calibrated or certified on their own.

  • The obstruction for mixtures of Gaussian copulas is correct and well illustrated (Figure 2 is enlightening).

  • Splits are paired across methods, code is released, and the appendices are thorough.

Critical weaknesses
  • Related work omits the GP view: the latent model is a finite mixture of Gaussian processes. Each component has mean μ_j(x) and kernel k_j(x, x′) = σ_j(x)σ_j(x′)[⟨ṽ_j(x), ṽ_j(x′)⟩ + 𝟙{x = x′}(1 − ‖ṽ_j(x)‖²)]. GMCP is therefore a copula of a mixture of warped GPs. The paper discusses neither mixtures of GP experts nor warped GPs.

  • " first copula process that is simultaneously conditioned on a context set, marginalization consistent, permutation equivariant, and able to express non-Gaussian dependence with a tractable likelihood" overlooks Student-t processes. These are consistent, non-Gaussian, and have tractable likelihoods, although they are elliptical.

Questions for authors

How does the method relate to a finite mixture of non-stationary (warped) Gaussian processes?

Reviewer_WdNqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper introduces a Gaussian mixture for the Copula density in the Gaussian copula neural process, which preserves the original properties of the Gaussian copula neural process (marginalisation consistency, etc.) while providing a mixture model extension to the basic Gaussian approach. The proposed method is similar in structure to the basc Gaussian conditional copula, trained with similar techniques and with embeddings etc. drawn from the current state of the art.

Critical strengths

This work does indeed generalize the prior work, replacing the Gaussian copula with a Gaussian mixture copula.

The benefit is demonstrated through synthetic experiments.

Critical weaknesses
  1. The advance proposed replaces the Gaussian copula in the existing Gaussian copula neural process with a Gaussian mixture copula. This is quite a straightforward modification, but the authors obscure the development (in my opinion) and could have stated the model much more simply and at an earlier stage.
  2. A significant part of the analysis either follows from the established properties of the copula neural process or comes from the existing papers as cited. The contributions are therefore more limited than stated.
  3. The experimental comparisons seem limited, applied to simple univariate (? Have I understood correctly?) continuous time models in the synthetic case.
Questions for authors

Why have you not included statistical significance tests to check if introducing the Gaussian mixture provides a sigifnicant improvement compared to the Gaussian one? An earlier Arxiv version seemed to contain these results, now deleted? Can you provide comparisons to other expressive baselines for the tasks, or at least explain why they cannot be included? I don’t think that the irregularity of time series is a key limitation, as many methods can still be applied by ignoring the spacing differences or adapting through resampling or interpolation, etc.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.